Chinese Text Classification via Bidirectional Lattice LSTM

In this paper, we investigate a bidirectional lattice LSTM (Bi-Lattice) network for Chinese text classification. The new network is different from the standard LSTM in adding shortcut paths which link the start and end characters of words, to control the information flow. Character-level features can flow into word-level by an extra gate, and word-level features are integrated into character-level via a weighted manner by another gate. Previous models take as input embeddings pre-trained by Skip-Gram model, we utilize word sememes in HowNet to further improve the word representation learning in our proposal. Our experiments show that Bi-Lattice gives better results compared with the state-of-the-art methods on two Chinese text classification benchmarks. Detailed analyses are conducted to show the success of our model in feature fusion, and the contribution of each component.